Merge pull request #507 from sonots/bp-480

[backport] Fix multi thread problem in PooledMemory

cupy/cuda/memory.pyx

@@ -419,44 +419,49 @@ class PooledMemory(Memory):
         self.size = chunk.size
         self.pool = pool
 
-    def __del__(self):
-        if self.ptr != 0:
-            self.free()
-
     def free(self):
         """Frees the memory buffer and returns it to the memory pool.
 
         This function actually does not free the buffer. It just returns the
         buffer to the memory pool for reuse.
 
         """
+        cdef Py_ssize_t ptr
+        ptr = self.ptr
+        if ptr == 0:
+            return
+        self.ptr = 0
         pool = self.pool()
-        if pool and self.ptr != 0:
-            hooks = memory_hook.get_memory_hooks()
-            if hooks:
-                device_id = self.device.id
-                pmem_id = id(self)
-                size = self.size
-                ptr = self.ptr
-                hooks_values = hooks.values()  # avoid six for performance
+        if pool is None:
+            return
+
+        hooks = None
+        # to avoid error at exit
+        mh = memory_hook
+        if mh is not None and mh.get_memory_hooks is not None:
+            hooks = mh.get_memory_hooks()
+        size = self.size
+        if hooks:
+            device_id = self.device.id
+            pmem_id = id(self)
+            hooks_values = hooks.values()  # avoid six for performance
+            for hook in hooks_values:
+                hook.free_preprocess(device_id=device_id,
+                                     mem_size=size,
+                                     mem_ptr=ptr,
+                                     pmem_id=pmem_id)
+            try:
+                pool.free(ptr, size)
+            finally:
                 for hook in hooks_values:
-                    hook.free_preprocess(device_id=device_id,
-                                         mem_size=size,
-                                         mem_ptr=ptr,
-                                         pmem_id=pmem_id)
-                try:
-                    pool.free(ptr, size)
-                finally:
-                    for hook in hooks_values:
-                        hook.free_postprocess(device_id=device_id,
-                                              mem_size=size,
-                                              mem_ptr=ptr,
-                                              pmem_id=pmem_id)
-            else:
-                pool.free(self.ptr, self.size)
-        self.ptr = 0
-        self.size = 0
-        self.device = None
+                    hook.free_postprocess(device_id=device_id,
+                                          mem_size=size,
+                                          mem_ptr=ptr,
+                                          pmem_id=pmem_id)
+        else:
+            pool.free(ptr, size)
+
+    __del__ = free
 
 
 cdef class SingleDeviceMemoryPool: